Proceedings of the 54th Hawaii International Conference on System Sciences | 2021

Communicating with Machines: Conversational Agents with Personality and the Role of Extraversion

Rangina Ahmad Dominik Siemon Susanne Robra-Bissantz TU Braunschweig TU Braunschweig TU Braunschweig [email protected] [email protected] [email protected]

Abstract environments [32]. With all these areas of application, Communication with conversational agents (CA) the design of CAs and specifically their effect on has become increasingly important. It therefore is human-computer interaction (HCI) has become even crucial to understand how individuals perceive more important. In prolonging casual conversations, interaction with CAs and how the personality of both the long-term goal is to take steps toward the AI dream the CA and the human can affect the interaction of general intelligence, that is, CAs that are able to experience. As personality differences are manifested provide casual opinions or therapeutic responses and in language cues, we investigate whether different have “a lot of personality” [3, p.3]. However, handling language style manifestations of extraversion lead to a open conversation effectively and providing the more anthropomorphized perception (specifically machine with the ability to converse with humans in a perceived humanness and social presence) of the natural way and in a manner that the user is satisfied or personality bots. We examine, whether individuals rate even enjoys the interaction, is still a major challenge communication satisfaction of a CA similar to their [22, 35, 59]. own personality as higher (law of attraction). The In order to have a better understanding of the nature results of our experiment indicate that highly and quality of human-machine interactions, researchers extraverted CAs are generally better received in terms have drawn from a variety of other disciplines, such as of social presence and communication satisfaction. psychology and sociology. In this context, many Further, incorporating personality into CAs increases studies have confirmed that findings in human-human perceived humanness. Although no significant effects interaction (HHI) can be applied to the interaction could be found in regard to the law of attraction, between machines and humans [36, 39, 41, 47]. For interesting findings about ambiverts could be made. instance, people treat a computer system as if it were a The outcomes of the experiment contribute towards human and thus project a certain level of designing personality-adaptive CAs. anthropomorphism upon the machine – this perceived humanness has been an important aspect for social 1. Introduction interactions [14, 37, 40, 52]. Also examined by many researchers is the construct of social presence, which describes the sense of connection that a user feels with In the past years, progress has been made in the their IT communication partner [12, 19, 52]. As a field of AI in recognizing and mimicking human matter of fact, humans will respond to computers and activities [22, 29]. Conversational Agents (CA) are to a CAs using the same elements of social interactions that certain degree capable of simulating such human they employ in HHI [37, 41]. In particular, researchers behavior, by interacting with their users via natural have shown that people respond to computer systems language – in written or spoken form [22, 35]. With the in similar ways to how they would respond to a human, introduction of Google Assistant, Apple's Siri or for instance by attributing certain personality traits to Amazon's Alexa, CAs have been made accessible to a computer partners [41, 52]. A specific set of wide range of users [35]. CAs have also been characteristics is on the one hand believed to explain incorporated in many companies, specifically on their the way people respond to others in social settings [47, websites and messenger platforms, for example to 57], and provides on the other hand an explanation assist customers during the sales process [31]. Another why it influences the quality of interactions [13, 47]. context, in which CAs play an increasingly important As for the question, which personality traits in specific role is in health and medical care, supporting are more desirable in CAs, there is no definite answer consumers with mental health challenges, or assisting to it. In fact, more and more researchers conclude that patients and elderly individuals in their living CAs should not have a static personality, but should

URI: https://hdl.handle.net/10125/71109 978-0-9981331-4-0 Page 4043 (CC BY-NC-ND 4.0) ideally adapt to the personality of the user [1, 2, 22, 53, framework of our conducted experiment. We present 55, 64]. As part of initial steps towards our research the results and discuss our research questions. goal of designing personality-adaptive CAs, we investigate existing theories, specifically social 2. Theory & Hypotheses presence and perceived humanness in connection with the personality trait extraversion. The extraversion/ introversion dichotomy has specifically received the Personality has been an essential topic in the most attention from researchers, as the underlying literature of HCI. Personality differences are components of extraversion have been well-established manifested in language and can thus be incorporated to date across various methodologies [7, 33]. Further, into intelligent systems such as CAs by using specific since incorporating different personality-based language cues that reflect a certain personality language styles increase social presence [42] and dimension. CAs can have (or simulate) a personality extraverted individuals are more likely to that can easily adapt according to a user’s personality anthropomorphize bots [47, 49], we pose the following to give the user the best possible interaction. In research question (RQ): research, the law of attraction is often used to determine which personality types and factors are RQ1: Are social presence and perceived preferred. However, an important prerequisite for this humanness increased when incorporating an is the general perception of a CA by its user and extraverted language style in human-machine specifically their perceived humanness and social conversation? presence

Another step towards a better understanding of how 2.1. Personality & Language Cues to increase interaction quality provides the law of attraction. According to this theory, humans prefer A well accepted theory of psychology is that human machines with a similar personality to their own [39, language reflects the emotional state and personality, 41]. On the one hand, studies confirm the law of based on the frequency with which certain categories attraction [39, 41], on the other hand however, there of words are used as well as the variations in word are some findings indicating people prefer interaction usage [7, 20, 63]. Personality is described as a based on opposing personality traits [27]. Since in this psychological construct that comprises someone’s case concepts of HHI cannot be clearly transferred to behaviors, emotions and cognitions derived from both HCI, and since there is a larger body of research biological and social factors [23, 47]. Theories of speaking for the law of attraction, we base our second psychology further assert that personality traits in RQ on this theory and investigate the following particular can be used to predict human emotions, question: cognitions and behaviors [42, 47]. In order to measure an individual’s personality, a widely used classification RQ2: Are extraverts perceiving higher of personality – the Big Five model – has been applied communication satisfaction when interacting with an in research [34]. For a comprehensive assessment of extraverted CA? individuals, the following five fundamental traits or dimensions have been defined and derived through Though researchers have given great thought to the factorial studies: Conscientiousness, openness, effects of different personality dimensions on HCI, yet, neuroticism, agreeableness and extraversion which only a few studies have dedicated their research on refers to the extent to which people enjoy company and personality – extraversion in particular – and their seek excitement and stimulation [10]. This specific set effect on perceived humanness and social presence. of characteristics is believed to explain the way people Further, as there are still disagreements regarding the respond to others in social settings, [47, 57] and also law of attraction and its opposing theory, we aim to provides an explanation as to why it influences the study whether different language style manifestations quality of interactions between people [13, 42, 47]. of extraversion (i.e. high, medium, low) have an effect Verbal interactions by means of language can therefore on users’ perceived communication satisfaction. be useful for capturing lower-level personality The paper is structured as follows: First, we processes, since language is more closely associated provide relevant theoretical background on personality- with objective behavioral outcomes than traditional based language cues, social presence and perceived personality measures [7]. Utterances for instance - humanness as well as on the law of attraction. We conveying a great deal of information about the derive our hypotheses, which are then tested in the speaker - comprise cues to the individual’s personality traits [33]. Both spoken language as well as written

Page 4044 language is unique from person to person, as humans applicable to HCI and consequently to CAs [12, 37]. express themselves verbally with their own distinctive According to the ‘computers are social actors’ (CASA) styles, even though the content of a message may be paradigm [37, 40, 41], humans will respond to the same [43]. By discovering correlations between a computers and CAs using the same elements of social range of linguistic variables and personality traits interactions that they employ in HHI [41]. In a variety across a wide range of linguistic levels, psychologists of studies, researchers have shown that people respond have documented the existence of such cues [33]. to computer systems in similar ways to how they Although research has found language markers for would respond to a human – for instance by attributing all Big Five traits, the extraversion/ introversion specific personality traits to computer partners [37, 52]. dichotomy has specifically received the most attention Relative to introverts, extraverts generally engage in from researchers, as the underlying components of more social activity, experience greater positive affect extraversion have been well-established to date across and well-being, and are reactive to external stimulation various methodologies [7, 33]. Most of the [15, 33, 50]. Further, an extraverted person will be personalities can be measured somewhere between two relatively gregarious and generally energized by extremities, since extraversion-introversion levels are external stimulus or people [16, 58]. An introvert on part of a single, continuous dimension of personality the contrary, will be comparatively less sociable and [56]. Table 1 shows a small overview of some of the more introspective [41, 54, 58]. These tendencies identified language cues for extraversion and indicate that humans may apply these social rules introversion, based on various studies by Scherer equally to HCI. (1979) [50], Furnham (1990) [15], Pennebaker and The construct of social presence can be used to King (1999) [43], Gill and Oberlander (2002) [18] and measure a user’s perception of a CA’s social Mairesse et al. (2007) [33]. characteristics. By incorporating a language style, that reflects a certain personality dimension, social 2.2. Social Presence & Perceived Humanness presence can be increased [44]. A high degree of social presence indicates enhanced trusting beliefs and Social presence refers to the subjective perceptions perceptions of enjoyment in the interaction with a CA of a medium that is used for communication, i.e. the [45]. We therefore hypothesize: sense of connection that a user feels with their communication partner [52]. It measures the extent to H1a: Social presence (SP) is perceived higher for which communication media has the ability to transmit extraverted than for introverted bots. social cues, such as transmitting sociable, personal or intimate aspects [40, 52]. While early research focused Perceived humanness is another relevant factor in on computer-mediated communication between the context of HCI. When interacting with an humans, recent studies state that social presence is also intelligent technological artifact such as a CA, the

Table 1. Language cues for extraversion and introversion Level Introvert Extravert Conversational Listen, less back-channel behavior Initiate conversation, more back-channel Behavior behavior Style Formal Informal Syntax Many nouns, adjectives, elaborated Many verbs, adverbs, pronouns constructions, many words per sentence, (implicit), few words per sentence, few many articles and negations articles, few negations Topic selection Self-focused, problem talk, dissatis- Pleasure talk, agreement, compliment, faction, single topic, few semantic errors many topics, many semantic errors Speech Slow speech rate, Many unfilled pauses, High speech rate, few unfilled pauses, long response latency, quiet, low voice short response latency, loud, high voice quality, low frequency variability quality, high frequency variability Lexicon Rich, high diversity, many exclusive and Poor, low diversity, few exclusive and inclusive words, few social words, few inclusive words, many social words, positive emotion words, many negative many positive emotion words, few emotion words negative emotion words

Page 4045 degree to which it is perceived as human has a positive focus (i.e., discussing one topic in depth) [15]. impact on aspects like trust [51] and service encounter Extraverts exert a more imprecise and “looser” style satisfaction [19]. Furthermore, a high degree of with reduced concreteness, whereas introverts exhibit a perceived humanness increases the degree of perceived more analytic, careful, precise and focused style [18]. competency in the CA [3], which in turn can improve Extraverts also use more positive emotion words and the overall interaction. Perceived humanness is also show more agreements and compliments than related to the willingness to disclose personal introverts [43]. We therefore make the following two information and the degree to which users feel more hypotheses: comfortable interacting with a CA [38]. Since people project a certain level of anthropomorphism upon the H2a: Extraverts assess their perceived computer, they treat the computer system as if it were a communication satisfaction for the extraverted bot human [37, 40, 52]. Specifically individuals who higher as for the introverted bot. scored high in extraversion have been found to be more likely to anthropomorphize a bot [47, 49]. This leads us H2b: Introverts assess their perceived to the following hypothesis regarding perceived communication satisfaction for the introverted bot humanness: higher as for the extraverted bot.

H1b: Perceived humanness (PH) is perceived Since not every individual falls into one of the two higher for extraverted than for introverted bots. extremes, but rather somewhere in the middle of the personality spectrum, a third group called ambiverts 2.3. Law of Attraction has been identified by psychologists [56]. As a solid mix of both extraversion and introversion, ambiverts Since personality may be one of the most important are balanced between the extremes [5]. As ambiverts factors in HCI among others (specifically when express traits from both personality styles and designing social CAs), studies have tried to define sometimes neither is dominant [21, 56], ambiverts are social CAs in order to determine their effect in often more flexible in their communication and human-machine interactions. A rich history of research interaction as they draw from a broader range of suggests that HCI largely parallel HHI [37, 39, 40, 46, communication options [14]. For instance, they engage 48], which is why researchers have examined machine in a flexible pattern of talking and listening [5]. personality based on human personality characteristics However, while ambiverts are able to adapt more [39, 46, 48]. The law of attraction (also called easily to the demands of a specific situation, they similarity-attraction theory) as one of the HHI theories sometimes have trouble determining which side of that can be transferred to HCI, posits that humans their personality to apply [56]. Based on these findings, prefer machines that have a similar personality to their we hypothesize as follows: own and are thus more comfortable interacting with them than with those CAs with an opposing personality H2c: Ambiverts do not assess communication [39, 41]. Users can better assess their counterparts satisfaction differently for extraverted or introverted when using similar personality traits, and information bots. has usually been rated as better and more trustworthy [64]. 3. Method As for the trait extraversion, Moon and Nass (1996) [36] found that dominant people prefer to interact with In order to address our research question and test an equally dominant chatbot. The same applies to our hypotheses, we conducted an online experiment. people with a submissive behavior [36]. Concerning We chose an experimental study to maximize internal verbal communication in HHI, it has been shown that validity [30] and to be able to validate if any observed speaker charisma for example strongly correlates with differences in communication satisfaction, perceived extraversion [33]. Relative to their introverted humanness and social presence, i.e. the dependent counterparts, extraverts tend to talk more, with fewer variables are caused by our manipulation, i.e. the pauses and hesitations, have shorter silences, a higher personality of our CAs (independent variable) [8]. Our verbal output and a less formal language, while experiment follows a within-subject design, in which introverts use a broader vocabulary [15, 18, 43, 50]. participants experience both conditions. The Research also showed that conversations between experiment took course over the span of two month extraverts are more expansive and characterized by a and had on average a completion time of 20 - 25 wider range of topics whereas a conversation between minutes. two introverts are more serious and have a greater topic

Page 4046 partners and to closely observe the dialogues with 3.1. Sample Raffi the CA. Subsequently, the subjects completed a survey including the established measuring construct A relatively large sample size of participants was communication satisfaction (CS) by Hecht (1978) [24], obtained to ensure more reliable results and greater in order to test H2a, H2b and H2c. Originally intended significance. 478 native English speaker or people with for HHI, we transferred this construct to a HCI context, English as their second language were recruited from since the inventory shows a high degree of reliability the crowdsourcing platform Mechanical Turk (mTurk) and validity when measuring communication as well as via personal networks. While the participants satisfaction with “actual and recalled conversations on mTurk were paid for the task, the people from our with another perceived to be a friend, acquaintance, or personal network were not compensated for their stranger” [17, p. 253]. The construct consists of 19 participation. Subjects of both sources did not differ items, and as suggested in the study, we used a 7-point regarding any relevant characteristics, such as their Likert scale. However, we adapted the phrasing of the demographic data or personality traits. Of the 478 items accordingly to our CA. For instance, we changed participants, we eliminated the data of 113 test persons the wording of the original item “The other person who cancelled the experiment in advance. Another 102 expressed a lot of interest in what I had to say” to subjects failed the attention checks built into the study “Raffi expressed a lot of interest in what I had to say”. or did not complete the personality test, leaving 263 In order to check hypotheses H1a and H1b, the participants (177 males, 84 females, 2 other). The age survey included the measures social presence (SP) and of the subjects ranged from 17 to 74 years, with an perceived humanness (PH). We used the perceived average age of 32.9 years. humanness measure (6 items) by Gill and Oberlander (2002) [17], which is measured on a 9-point Likert 3.2. Procedure scale (1 = low, 9 = high) and social presence (5 items) by Gefen and Straub (2004) [17], which was measured In the first step of the experiment, the subjects were on a 7-point Likert scale (1 = low, 7 = high). The asked to conduct an online personality test based on survey further covered demographic questions (gender, Johnson's (2014) [28] 120-item IPIP NEO-PI-R. Their age, language) and an open question (Do you think that results were values between 0-120, with 0 being a low the concept of a personality adaptive CA can enhance a score and 120 a high score in a specific personality person's interaction experience?). To check if our dimension. The unique identification numbers the manipulation was successful and the participants participants received after finishing their personality perceived the ExtraBot as more extraverted as the tests were also used throughout the second part of the IntroBot, we asked the participants to indicate the experiment in order to provide anonymity between the extent to which the attributes sociable, talkative, active, experimenters and the subjects. We chose a within- impulsive, outgoing, shy, reticent, passive, deliberate, subject design, where the participants were exposed to reserved apply to the CAs [4]. While the first five both levels of treatment one after the other [9], i.e. the items reflect high extraversion, the last five attributes subjects were randomly assigned to LimeSurvey (an stand for low extraversion [4]. The measure is online survey tool), where they either watched a video summarized as the construct extraversion (EX). of a conversation between an extraverted Chatbot (ExtraBot) and a fictional human first and a 3.3. Conversational Agent Design conversation of the introverted Chatbot (IntroBot) second or vice versa. This way we ensured that We used the conversational design tool Botsociety individual differences were not distorting the results, [6] to visualize our CAs. The two pre-defined dialogue since every subject acted as their own control. This structures are communications between the CA Raffi reduced the chance of confounding factors. The order and the humans Jamie and Francis, respectively. While of the two conditions was hence distributed randomly, Jamie and the ExtraBot are intended to take on an and the dependent variable was measured after each extraverted personality, Francis and the IntroBot are condition by means of a subsequent survey. Every both of introverted nature. We achieved this by participant was provided with the exact same sets of adapting their respective language style according to information for the experiment [11]. the personality dimensions, specifically by using the The videos lasted about 3 minutes, skipping was language cues for extraversion and introversion listed not allowed, and a control question was added at the in Table 1. In order to find out whether the language end of the video. The participants’ task was to put styles used in the conversations truly reflected the themselves in the shoes of the human conversation dimensions extraversion and introversion, we used IBM Watson’s Personality Insights tool [26] for

Page 4047 verification purposes. Based on text that is being 3.4. Results analyzed, the personality mining service returns different percentiles for the Big Five dimensions. We used confirmatory factor analysis to test our Percentiles are defined as scores that compare one measures CS, PH, SP and EX (manipulation check). person to a broader population [26]. While the After dropping three items of CS and three items of EX ExtraBot dialogue received a score of 83%, meaning (manipulation check), all items had significant positive that the ExtraBot is more extraverted than 83% of the factor loadings greater than .600. Furthermore, we people in the population, the IntroBot had a percentile calculated composite reliability (CR) and Cronbach’s of 36%, thus scoring low in extraversion (and high in alpha, both indicating reliable factors (see Table 2) introversion). With these results, we incorporated the [60]. two language styles into our CA design. The topics of All analyses were carried out using the statistical the dialogues were everyday conversations, such as computing software RStudio (Version 1.2.5033). Due weekend plans, music, books and travel. However, to the use of ordinal scales [62], a non-normality compared to the ExtraBot, the IntroBot sticks mainly to assumption and a within-subject design (paired two topics (travel and books) and has a rather rich sample), we chose the nonparametric Wilcoxon signed- vocabulary throughout the whole dialogue by using rank test to test for significant differences (CI = .95). many words per sentence. The IntroBot also uses fewer Table 2 provides the descriptive statistics, Cronbach’s emotional words and has fewer semantic errors (“At alphas and the composite reliability of the measures. least it's Friday! How was your day?”) compared to the ExtraBot. The ExtraBot on the other hand talks about Table 2. Descriptive statistics, Cronbach’s alpha many topics in a short amount of time, uses a rather and composite reliability informal language (e.g. “what r u up to?”, “…cause Measure MEB SDEB MIB SDIB α CR TGIF!”), compliments and uses many positive emotion CS 5.25 0.98 5.08 0.94 .90 .89 words (“Sounds amazing!” “Have fun at the party!”) PH 7.22 1.35 7.08 1.48 .91 .91 and few words per sentence (“Nope. Locals as well.”, SP 5.50 1.05 5.30 1.19 .89 .89 “Told ya!”). EX 4.68 0.86 4.11 0.89 .81 .73 The videos of the complete conversations can be watched at the following links: After the experiment, the participants indicated the https://youtu.be/B1N7XwcdCE0, https://youtu.be/d26eKdHBKeQ. Figure 1 shows a extent to which the attributes concerning extraversion snippet of the two conversations between the ExtraBot apply to the ExtraBot and IntroBot. The conducted and Jamie (left) and the IntroBot and Francis (right). Wilcoxon test revealed, that there is a significant difference (W = 46374, p < .001) in the rated attributes of extraversion between the ExtraBot and the IntroBot i.e. the ExtraBot was perceived as more extraverted than the IntroBot. Since our manipulation check was successful, we continued testing our hypotheses. Social presence was rated higher for the ExtraBot (M = 5.50) than for the IntroBot (M= 5.30), which is a significant difference (W = 37826, p = .048), supporting H1a. perceived humanness was rated slightly higher for the ExtraBot (M = 7.22) than for the IntroBot (M = 7.08). However, the difference is not significant (W = 36981, p = .169), which is why H1b cannot be supported. In order to test H2a, H2b and H2c, the population was divided into three different groups, representing participants with a low extraversion (introverts), medium extraversion (ambiverts) and high extraversion (extraverts). Clustering was performed by means of the R package Ckmeans.1d.dp, a procedure for optimal k- means clustering in one dimension [61]. This resulted Figure 1. Mockups of the ExtraBot (left) and in three groups with n = 58 for introverts (Cluster IntroBot (right) Conversation center = 55.20, Min = 30, Max = 66), n = 114 for ambiverts (Cluster center = 77.81, Min = 67, Max = 85) and n = 91 for extraverts (Cluster

Page 4048 center = 93.20, Min = 86, Max = 111). Table 3 shows introverts (n = 58). The results indicate that all three the descriptive statistic of each cluster, for groups rated CS higher for the ExtraBot, even though communication satisfaction and the results of the no significant difference was found. So while Wilcoxon tests. hypotheses H1a and H1b cannot be supported, it still can be said that overall the participants rated CS higher Table 3. Communication satisfaction by clusters for the ExtraBot (M = 5.25) than for the IntroBot Introverts Ambiverts Extraverts (M = 5.08), which is a significant difference n 58 114 91 (W = 38057, p = .046). To our surprise, introverts Communication satisfaction rated the CS of the ExtraBot similar high as the MEB 5.40 5.03 5.42 extraverts, thus not confirming the law of attraction. A MIB 5.30 4.84 5.20 possible explanation for this may be that the Wilcoxon W = 4380 W = 7152 W = 2036 introverted participants were more drawn to an test p = 0.501 p = 0.189 p = 0.113 extraverted CA, because the ExtraBot’s SP was significantly higher than the IntroBot’s SP. And as The results of the Wilcoxon tests show no studies have shown, the higher the degree of social significant difference in communication satisfaction presence, the more trust beliefs and perceptions of between the ExtraBot and IntroBot within the clusters. enjoyment in the interaction with a CA a user has. As Therefore, only H2c is supported, whereas we reject assumed for the ambiverts, the results show that they H2a and H2b. rated CS in both CAs as the lowest. H2c thus can be supported, as they expressed no clear preference for any of the two bots. 4. Discussion In order to find a clearer answer for RQ2, we took a closer look at the dataset, particularly the data of The mean values of the PH measure (PHEB = 7.22; participants who scored either very high or very low in PHIB = 7.08) are above the mean of the scale (4.50) for extraversion. The tendencies towards their bot both the ExtraBot and the IntroBot. These substantially preferences were mixed: We found extraverts, who increased values demonstrate that the incorporation of clearly rated the CS for the ExtraBot as much higher as language cues and the associated depiction of a for the IntroBot (confirming the law of attraction), but personality has led to a human perception of the CAs, then again there were also participants high in even though there is no significant difference between extraversion, who indicated that they perceived a the two CAs. The SP scale also clearly shows an higher CS for the IntroBot. The same could be increased value for both CAs (SPEB = 5.50; SPIB = 5.30; observed vice versa for introverts. The rather high Mean of the scale = 3.50). The Wilcoxon test number of outliers, who preferred the CA with the furthermore reveals a significant difference between opposing personality trait explain why there could not the ExtraBot and the IntroBot for SP. SP was rated be found a significant effect on CS. We further believe, higher for the ExtraBot than for the IntroBot, a direct interaction between the participants and the indicating that the ExtraBot’s ability to transmit social CAs - rather than letting the subjects evaluate the cues, its personality and sociability with the conversation of two fictional characters and assuming incorporation of language cues was better than the all participants were able to emphasize with Jamie and IntroBot. Concerning RQ1, both SP and PH were Francis - may have had a significant effect on CS. So increased by personality differences manifested in the question, whether extraverts perceive higher CS language use. However, while an extraverted language when interacting with an extraverted CA cannot be style as opposing to an introverted language use led to answered with a definite yes. a significantly higher SP, the same cannot be said for Though 3 out of 5 hypotheses were rejected - in PH. The subjects did not perceive the ExtraBot as regard to our research project towards designing significantly more human than the introverted CA. personality-adaptive CAs, we take the following Nevertheless, the results show that the use of aspects from this experiment: First, incorporating personality-based language styles in HCI can achieve personality that is manifested in a specific language an increased SP as well as PH. These theories in turn style (in this case notably extraversion), increases a are fundamental aspects for an enhanced interaction bot’s social presence. Second, using language cues that quality. are specific for a certain personality dimension Clustering the participants according to their level (whether high or low extraversion), lead to the CA of extraversion, resulted in a large number of ambiverts being perceived more as a human. Third, as a high (n = 114), followed by the second biggest group of number of the population is believed to be ambiverted, extraverts (n = 91) and the smallest group consisting of rather than extremely introverted or extraverted [56]

Page 4049 (matching our datasat), further research has to be done [3] Araujo, T., “Living up to the chatbot hype: The influence on how to enhance ambiverts’ interaction quality. In of anthropomorphic design cues and communicative fact, since ambiverts seem to have no specific agency framing on conversational agent and company preferences concerning their communication style and perceptions”, Computers in Human Behavior 85, 2018, pp. 183–189. rather adapt according to a given situation, it is the more important that a CA itself should be able to [4] Back, M.D., S.C. Schmukle, and B. Egloff, “Predicting switch and adapt to the preferred communication style actual behavior from the explicit and implicit self- of its user. Fourth, the results of our experiment clearly concept of personality.”, Journal of personality and demonstrated that the law of attraction cannot always social psychology 97(3), 2009, pp. 533. be confirmed. The number of participants who showed [5] Bernstein, E., “Not introvert, nor extrovert: The adaptable a preference for a CA with an opposing personality ambivert.”, Wall Street Journal, 2015, pp. 1–3. trait were too high to ignore and label them just as “a [6] Botsociety, “Design, preview and prototype your next few outliers”. This again argues for the fact that CA chatbot or voice assistant”, 2020. https://botsociety.io personality should not be designed statically but dynamically, so that it easily adapts to user personality [7] Boyd, R.L., and J.W. Pennebaker, “Language-based - regardless of whether the user is highly extraverted, personality: a new approach to personality in a digital world”, Current Opinion in Behavioral Sciences 18, introverted or has any other dominant Big Five 2017, pp. 63–68. personality trait. It can be concluded that a “one-size- fits-all”- design for CAs does not necessarily enhance [8] Brewer, M.B., and W.D. Crano, “Research design and interaction quality, since each user has unique issues of validity”, Handbook of research methods in personality traits, abilities, perceptual preferences, social and personality psychology, 2000, pp. 3–16. experiences, etc., that directly affect every interaction [9] Charness, G., U. Gneezy, and M.A. Kuhn, “Experimental process. Thus, communicating with machines should methods: Between-subject and within-subject design”, be personalized. Journal of Economic Behavior & Organization 81(1), 2012, pp. 1–8. 5. Conclusion [10] Costa, P.T., and R.R. McCrae, “The revised neo personality inventory (neo-pi-r)”, The SAGE handbook of personality theory and assessment 2(2), 2008, pp. As an initial step towards our research goal of 179–198. designing personality-adaptive CAs, we investigated existing theories, such as social presence, perceived [11] Dennis, A.R., and J.S. Valacich, “Conducting humanness and the law of attraction in connection with experimental research in information systems”, Communications of the association for information the personality trait extraversion. Our experiment systems 7(1), 2001, pp. 5. demonstrated that incorporating personality into a bot, makes the CA being perceived as more human. As [12] Diederich, S., M. Janssen-Müller, A.B. Brendel, and S. opposing to the introverted bot, the extraverted CA in Morana, “Emulating Empathetic Behavior in Online particular was better received in terms of social Service Encounters with Sentiment-Adaptive Responses: Insights from an Experiment with a presence. Though communication satisfaction of the Conversational Agent”, ICIS 2019 Proceedings, 2019. extraverted CA was slightly rated higher, no significant effects could be found in regard to the law of [13] Driskell, J.E., G.F. Goodwin, E. Salas, and P.G. O’Shea, attraction. Interesting findings about ambiverts could “What makes a good team player? Personality and be made, however more research is required. Our team effectiveness.”, Group Dynamics: Theory, findings contribute to CA design and provide insights Research, and Practice 10(4), 2006, pp. 249. for researchers who want to follow up with research on [14] Fink, J., “Anthropomorphism and Human Likeness in personality in CAs. the Design of Robots and Human-Robot Interaction”, Social Robotics, Springer (2012), 199–208. 6. References [15] Furnham, A., “Faking personality questionnaires: Fabricating different profiles for different purposes”, Current psychology 9(1), 1990, pp. 46–55. [1] Ahmad, R., D. Siemon, D. Fernau, and S. Robra-Bissantz, [16] Furnham, A., and C.R. Brewin, “Personality and “Introducing" Raffi": A Personality Adaptive happiness”, Personality and individual differences Conversational Agent.”, PACIS 2020, 28. 11(10), 1990, pp. 1093–1096. [2] Ahmad, R., D. Siemon, and S. Robra-Bissantz, “ExtraBot [17] Gefen, D., and D.W. Straub, “Consumer trust in B2C e- vs IntroBot: The Influence of Linguistic Cues on Commerce and the importance of social presence: Communication Satisfaction”, AMCIS 2020.

Page 4050 experiments in e-Products and e-Services”, Omega traits”, ACM Transactions on Interactive Intelligent 32(6), 2004, pp. 407–424. Systems (TiiS) 6(4), 2016, pp. 28. [18] Gill, A.J., and J. Oberlander, “Taking care of the [32] Laranjo, L., A.G. Dunn, H.L. Tong, et al., linguistic features of extraversion”, Proceedings of the “Conversational agents in healthcare: a systematic Annual Meeting of the Cognitive Science Society, review”, Journal of the American Medical Informatics (2002). Association 25(9), 2018, pp. 1248–1258. [19] Gnewuch, U., S. Morana, and A. Maedche, “Towards [33] Mairesse, F., M.A. Walker, M.R. Mehl, and R.K. Designing Cooperative and Social Conversational Moore, “Using Linguistic Cues for the Automatic Agents for Customer Service”, ICIS, (2017). Recognition of Personality in Conversation and Text”, Journal of Artificial Intelligence Research 30, 2007, [20] Golbeck, J., C. Robles, M. Edmondson, and K. Turner, pp. 457–500. “Predicting personality from twitter”, Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE [34] McCrae, R.R., and O.P. John, “An introduction to the Third Inernational Conference on Social Computing five-factor model and its applications”, Journal of (SocialCom), 2011 IEEE Third International personality 60(2), 1992, pp. 175–215. Conference on, IEEE (2011), 149–156. [35] McTear, M., Z. Callejas, and D. Griol, The [21] Grant, A.M., “Rethinking the extraverted sales ideal: conversational interface: Talking to smart devices, The ambivert advantage”, Psychological science 24(6), Springer, 2016. 2013, pp. 1024–1030. [36] Moon, Y., and C. Nass, “How ‘real’ are computer [22] Grudin, J., and R. Jacques, “Chatbots, Humbots, and the personalities? Psychological responses to personality Quest for Artificial General Intelligence”, Proceedings types in human-computer interaction”, Communication of the 2019 CHI Conference on Human Factors in research 23(6), 1996, pp. 651–674. Computing Systems - CHI ’19, ACM Press (2019), 1– 11. [37] Nass, C., and Y. Moon, “Machines and Mindlessness: Social Responses to Computers”, Journal of Social [23] Hall, C.S., G. Lindzey, and J.B. Campbell, Theories of Issues 56(1), 2000, pp. 81–103. personality, Wiley New York, 1957. [38] Nass, C., and Y. Moon, “Machines and mindlessness: [24] Hecht, M.L., “The Conceptualization and Measurement Social responses to computers”, Journal of social of Interpersonal Communication Satisfaction”, Human issues 56(1), 2000, pp. 81–103. Communication Research 4(3), 1978, pp. 253–264. [39] Nass, C., Y. Moon, B.J. Fogg, B. Reeves, and D.C. [25] Holtgraves, T., and T.-L. Han, “A procedure for Dryer, “Can computer personalities be human studying online conversational processing using a chat personalities?”, International Journal of Human- bot”, Behavior research methods 39(1), 2007, pp. 156– Computer Studies 43(2), 1995, pp. 223–239. 163. [40] Nass, C., J. Steuer, and E.R. Tauber, “Computers are [26] IBM Watson PI, “IBM Watson Personality Insights”, social actors”, Proceedings of the SIGCHI conference 2020. https://personality-insights-demo.ng.bluemix.net/ on Human factors in computing systems, (1994), 72– 78. [27] Isbister, K., and C. Nass, “Consistency of personality in interactive characters: verbal cues, non-verbal cues, [41] Park, E., D. Jin, and A.P. del Pobil, “The law of and user characteristics”, International Journal of attraction in human-robot interaction”, International Human-Computer Studies 53(2), 2000, pp. 251–267. Journal of Advanced Robotic Systems 9(2), 2012, pp. 35. [28] Johnson, J.A., “Measuring thirty facets of the Five Factor Model with a 120-item public domain inventory: [42] Peeters, M.A., H.F. Van Tuijl, C.G. Rutte, and I.M. Development of the IPIP-NEO-120”, Journal of Reymen, “Personality and team performance: a meta- Research in Personality 51, 2014, pp. 78–89. analysis”, European Journal of Personality: Published for the European Association of Personality [29] Kaplan, A., and M. Haenlein, “Siri, Siri, in my hand: Psychology 20(5), 2006, pp. 377–396. Who’s the fairest in the land? On the interpretations, illustrations, and implications of artificial intelligence”, [43] Pennebaker, J.W., and L.A. King, “Linguistic styles: Business Horizons 62(1), 2019, pp. 15–25. Language use as an individual difference.”, Journal of personality and social psychology 77(6), 1999, pp. [30] Karahanna, E., I. Benbasat, R. Bapna, and A. Rai, 1296. “Editor’s comments: opportunities and challenges for different types of online experiments”, MIS Quarterly [44] Qiu, L., and I. Benbasat, “Evaluating anthropomorphic 42(4), 2018, pp. iii–x. product recommendation agents: A social relationship perspective to designing information systems”, Journal [31] Knijnenburg, B.P., and M.C. Willemsen, “Inferring of Management Information Systems 25(4), 2009, pp. capabilities of intelligent agents from their external 145–182.

Page 4051 [45] Qiu, L., and I. Benbasat, “Evaluating anthropomorphic personality and social behavior”, Journal of personality product recommendation agents: A social relationship 51(3), 1983, pp. 497–516. perspective to designing information systems”, Journal of management information systems 25(4), 2009, pp. [55] Strohmann, T., D. Siemon, and S. Robra-Bissantz, 145–182. “Designing Virtual In-vehicle Assistants: Design Guidelines for Creating a Convincing User [46] Robert, L., “Personality in the human robot interaction Experience”, AIS Transactions on Human-Computer literature: A review and brief critique”, Robert, LP Interaction 11(2), 2019, pp. 54–78. (2018). Personality in the Human Robot Interaction Literature: A Review and Brief Critique, Proceedings [56] Taylor, M., and others, “Personality Styles: Why They of the 24th Americas Conference on Information Matter in the Workplace”, Economic Alternatives(1), Systems, Aug, (2018), 16–18. 2020, pp. 148–163. [47] Robert, L., R. Alahmad, C. Esterwood, S. Kim, S. You, [57] Thoresen, J.C., Q.C. Vuong, and A.P. Atkinson, “First and Q. Zhang, “A Review of Personality in Human– impressions: Gait cues drive reliable trait judgements”, Robot Interactions”, Available at SSRN 3528496, 2020. Cognition 124(3), 2012, pp. 261–271. [48] Robert, L.P., R. Alahmad, C. Esterwood, S. Kim, S. [58] Thorne, A., “The press of personality: A study of You, and Q. Zhang, “A Review of Personality in conversations between introverts and extraverts.”, Human Robot Interactions”, arXiv preprint Journal of Personality and Social Psychology 53(4), arXiv:2001.11777, 2020. 1987, pp. 718. [49] Salem, M., G. Lakatos, F. Amirabdollahian, and K. [59] Turing, A.M., “Computing Machinery and Intelligence”, Dautenhahn, “Would you trust a (faulty) robot? Effects Mind, New Series 59(236), 1950, pp. 433–460. of error, task type and personality on human-robot [60] Urbach, N., F. Ahlemann, and others, “Structural cooperation and trust”, 2015 10th ACM/IEEE equation modeling in information systems research International Conference on Human-Robot Interaction using partial least squares”, Journal of Information (HRI), IEEE (2015), 1–8. technology theory and application 11(2), 2010, pp. 5– [50] Scherer, K.R., Personality markers in speech, 40. Cambridge University Press, 1979. [61] Wang, H., and M. Song, “Ckmeans. 1d. dp: optimal k- [51] Schroeder, J., and M. Schroeder, “Trusting in Machines: means clustering in one dimension by dynamic How Mode of Interaction Affects Willingness to Share programming”, The R journal 3(2), 2011, pp. 29. Personal Information with Machines”, Proceedings of [62] Wu, H., and S.-O. Leung, “Can Likert Scales be Treated the 51st Hawaii International Conference on System as Interval Scales?—A Simulation Study”, Journal of Sciences, (2018). Social Service Research 43(4), 2017, pp. 527–532. [52] Schuetzler, R., G. Grimes, J. Giboney, and J. [63] Yarkoni, T., “Personality in 100,000 words: A large- Nunamaker, “The Influence of Conversational Agents scale analysis of personality and word use among on Socially Desirable Responding”, Proceedings of the bloggers”, Journal of research in personality 44(3), 51st Hawaii International Conference on System 2010, pp. 363–373. Sciences, 2018, pp. 283–292. [64] Zumstein, D., and S. Hundertmark, “Chatbots - An [53] Smestad, T.L., and F. Volden, “Chatbot Personalities Interactive Technology for Personalized Matters”, Internet Science, Springer International Communication, Transactions and Services.”, IADIS Publishing (2019), 170–181. International Journal on WWW/Internet 15(1), 2017. [54] Snyder, M., “The influence of individuals on situations: Implications for understanding the links between

Page 4052